Deep neural networks (DNNs) continue to demonstrate superior generalization performance in an increasing range of applications, including speech recognition and image understanding. Recent innovations in compression algorithms, design of efficient architectures and hardware accelerators have prompted a rapid growth in deploying DNNs on mobile and IoT devices to redefine user experiences. Relying on the superior inference quality of DNNs, various voice-enabled devices have started to pervade our everyday lives and are increasingly used for, e.g., opening and closing doors, starting or stopping washing machines, ordering products online, and authenticating monetary transactions. As the popularity of these voice-enabled services increases, so does their risk of being attacked. Recently, DNNs have been shown to be extremely brittle under adversarial attacks and people with malicious intentions can potentially exploit this vulnerability to compromise DNN-based voice-enabled systems. Although some existing work already highlights the vulnerability of audio models, very little is known of the behaviour of compressed on-device audio models under adversarial attacks. This paper bridges this gap by investigating thoroughly the vulnerabilities of compressed audio DNNs and makes a stride towards making compressed models robust. In particular, we propose a stochastic compression technique that generates compressed models with greater robustness to adversarial attacks. We present an extensive set of evaluations on adversarial vulnerability and robustness of DNNs in two diverse audio recognition tasks, while considering two popular attack algorithms: FGSM and PGD. We found that error rates of conventionally trained audio DNNs under attack can be as high as 100%. Under both white- and black-box attacks, our proposed approach is found to decrease the error rate of DNNs under attack by a large margin.

https://dl.acm.org/doi/abs/10.1145/3397332